Data Descriptives

Motion

We start with n= 74. In these analyses, I used a stringent pipeline, censoring and interpolating over vols with > FD of 0.5 mm or DVARS > 1.75. Those are excluded from the timeseries used to calculate the networks used here. I excluded anyone with > 50% frames censored (following Yeo, 2011) or mean motion > 1 mm, or max motion over 10 mm, leaving us with 70 participants, 26 of which have a second run.

The dataframe below shows all runs for subjects who meet this criteria. At the moment, I’m just examining the first run, not averaging in those who have a second run.

Data Frame Summary

motion

Dimensions: 92 x 7
Duplicates: 0
No Variable Stats / Values Freqs (% of Valid) Valid Missing
1 age_scan [numeric] Mean (sd) : 7 (1.3) min < med < max: 4.1 < 7 < 10.6 IQR (CV) : 2 (0.2) 68 distinct values 92 (100%) 0 (0%)
2 run [factor] 1. run-01 2. run-02 3. run-03
65(70.7%)
26(28.3%)
1(1.1%)
92 (100%) 0 (0%)
3 fd_mean [numeric] Mean (sd) : 0.3 (0.2) min < med < max: 0 < 0.2 < 0.8 IQR (CV) : 0.2 (0.6) 92 distinct values 92 (100%) 0 (0%)
4 pctSpikesFD [numeric] Mean (sd) : 0.1 (0.1) min < med < max: 0 < 0.1 < 0.4 IQR (CV) : 0.1 (1) 69 distinct values 92 (100%) 0 (0%)
5 relMeanRMSMotion [numeric] Mean (sd) : 0.3 (0.2) min < med < max: 0 < 0.2 < 0.8 IQR (CV) : 0.2 (0.6) 92 distinct values 92 (100%) 0 (0%)
6 nSpikesDV [integer] Mean (sd) : 11 (9.2) min < med < max: 0 < 9 < 35 IQR (CV) : 15 (0.8) 30 distinct values 92 (100%) 0 (0%)
7 relMaxRMSMotion [numeric] Mean (sd) : 2.7 (2.3) min < med < max: 0.1 < 2 < 9.6 IQR (CV) : 3 (0.9) 92 distinct values 92 (100%) 0 (0%)

Generated by summarytools 0.9.4 (R version 3.6.1)
2019-08-28

SES

Data Frame Summary

ses

Dimensions: 70 x 9
Duplicates: 6
No Variable Stats / Values Freqs (% of Valid) Valid Missing
1 male [factor] 1. Female 2. Male
44(62.9%)
26(37.1%)
70 (100%) 0 (0%)
2 race2 [factor] 1. White 2. Black 3. Other
16(23.5%)
39(57.4%)
13(19.1%)
68 (97.14%) 2 (2.86%)
3 ethnicity [factor] 1. Not Hispanic or Latino 2. Hispanic or Latino
60(85.7%)
10(14.3%)
70 (100%) 0 (0%)
4 has_diagnoses [integer] Min : 0 Mean : 0 Max : 1
0:68(97.1%)
1:2(2.9%)
70 (100%) 0 (0%)
5 parent1_edu [integer] Mean (sd) : 14.9 (2.7) min < med < max: 12 < 14 < 20 IQR (CV) : 6 (0.2)
12:26(37.7%)
14:10(14.5%)
16:13(18.8%)
18:16(23.2%)
20:4(5.8%)
69 (98.57%) 1 (1.43%)
6 parent2_edu [integer] Mean (sd) : 13.5 (4.4) min < med < max: 0 < 12 < 20 IQR (CV) : 4 (0.3)
0:3(5.1%)
10:4(6.8%)
12:27(45.8%)
14:4(6.8%)
16:8(13.6%)
18:7(11.9%)
20:6(10.2%)
59 (84.29%) 11 (15.71%)
7 income_median [integer] Mean (sd) : 62761.2 (55231.1) min < med < max: 2500 < 42500 < 2e+05 IQR (CV) : 67000 (0.9) 11 distinct values 67 (95.71%) 3 (4.29%)
8 monthslive_iflostincome [integer] Mean (sd) : 2.7 (1.5) min < med < max: 1 < 3 < 5 IQR (CV) : 3 (0.5)
1:20(30.3%)
2:11(16.7%)
3:16(24.2%)
4:7(10.6%)
5:12(18.2%)
66 (94.29%) 4 (5.71%)
9 childaces_sum_ignorenan [integer] Mean (sd) : 1.1 (1.4) min < med < max: 0 < 1 < 6 IQR (CV) : 2 (1.2)
0:32(45.7%)
1:16(22.9%)
2:11(15.7%)
3:7(10.0%)
4:2(2.9%)
5:1(1.4%)
6:1(1.4%)
70 (100%) 0 (0%)

Generated by summarytools 0.9.4 (R version 3.6.1)
2019-08-28

Cognition-NA for now

WPPSI and WISC scores are not totally finalized, so decided not to look at cognition for the moment, unless environmental effects are not interesting. For the future.

Age and measures of network segregation

We find that overall, all measures of network segregation significantly increase with age in our dataset, with remarkable consistency. We examined within-network connectivity, between network connectivity, system segregation (as calculated in Chan et al. 2018), the modul_avgarity quality index, and the participation coefficient (summed across negative and positive weights).

We controlled for sex, mean framewise displacement, percent of FD spikes in the data, the number of volumes a participant had, and the average weight of the functional network.

The number of communities detected using modul_avgarity maximization on this data does not significantly decrease with age (each subject ran 100x, averaged Q and k, range in k is 2.1, 4.78). ###Non-linear effects of age?

We also used restricted least ratio tests to test for the presence of non-linearity in our data. As you can see in the plots below, which use GAMs, most measures are linear and look no different from the linear models above.

Tests of non-linearity confirmed that no non-linear effects are present. The wiggly participation coefficient line is just over-fitting.

Which systems are driving this effect?

We fit the same model, controlling for sex, mean framewise displacement, percent of FD spikes in the data, the number of volumes a participant had, and the average weight of the functional network, to between- and within-system connectivity for each of the Yeo 7 systems.

We find the strongest effects are in the default mode system, between default and attentional networks.

None of these showed significantly non-linear effects, either.

Effects on default mode connectivity survive FDR correction across the tests conducted, with DMN-VAN p_fdr=0.0066791 and DMN-DAN p_fdr= 0.0076409. Vis-DAN p_fdr= 0.0770316.

Environment and Measures of Network Segregation

In summary, there are no main effects of SES, race, or ACES sum on any of the measures of network segregation. For examining results with the SES composite, I controlled for race (3-category) and ethnicity. For future reference, the estimates and p-values for other covariates (excluding the intercept) don’t change based on the reference level of a factor (re: Allyson’s concerns in Black vs. White ref category).

Using the same models as above, age_scan + male + fd_mean + avgweight + pctSpikesFD + size_t, we see no significant associations between any of the measures of network segregation and either parent 1 education, family income (median of bin), or an SES composite of the two. Race is also not significantly associated with any of the measures of network segregation. I also looked at the sum of child ACES, which doesn’t show any significant associations in the model above.

However, there are effects of the binned ACES (into 0-1, 2, or 3+) on within- and between-network connectivity as well as the participation coefficient. We see that 2 ACES is associated with significantly less within- and more between-network connectivity (p=3.311402710^{-10}), and a higher average participation coefficient (p=0.732495), lower system segregation (p=0.0325336).

Interactions between environment and age

No significant interactions between any of the environmental variables above and age. However, age x income is marginally predictive of within- and between-network connectivity.

p=0.0941697 for within-network connectivity and p=0.0941697 for between-network connectivity. However, this does not seem consistent across different SES metrics (education vs. income, etc.)

System-specific connectivity and environment main effects

I decided to look specifically at the networks that show strong age effects in our age range, that is, visual to dorsal attention and DMN to attentional networks, with the idea that previously the systems that showed stronger age effects also showed environmental interactions.

Visual to DAN

  • Significant + effect of parent education on avg connectivity (p = 0.01), not significant when controlling for race and ethnicity.

DMN to DAN

  • Significant - effect of ses composite on avg connectivity (p =0.02), not significant when controlling for race and ethnicity.
  • Significant - effect of parent education on avg connectivity (p =0.02), not significant when controlling for race and ethnicity.
  • Significant - effect of median income on avg connectivity (p =0.03), not significant when controlling for race and ethnicity.

DMN to VAN

  • Nothing

All systems connectivity and environment main effects

As Allyson suggested on 8/6, I ran models looking for a main effect of environment across all networks and fdr-corrected. I chose the SES composite for this analysis, since that seemed the most principled. No networks showed significant main effects of the environment after FDR correction when not controlling for race and ethnicity. When controlling for race and ethnicity, networks that showed significant effects of the environment were MOT-FPN, p_FDR=0.0384113 and VAN-LIM p_fdr=0.0922318.

Three-category ACES also does not show a significant environmental main effect after FDR-correction.

System-specific connectivity and environment interactions

These results are all controlling for race.

Visual to DAN

  • Significant interaction between age and ses composite on avg connectivity (p = 0.001)
  • Significant interaction between age and income on avg connectivity (p = 0.002)

DMN to DAN

  • Significant interaction between age and parent ed on avg connectivity (p = 0.041)

DMN to VAN

  • Significant interaction between age and ses composite on avg connectivity (p = 0.04)
  • Marginal interaction between age and income on avg connectivity (p = 0.06)
  • Marginal interaction between age and parent ed on avg connectivity (p = 0.09)

As suggested by Allyson on 8/6, I also ran some models looking at whether VIS-VIS, DMN-DMN, or FPN-DMN show age x SES interactions.

Vis to Vis

  • Nothing

DMN to DMN

  • Nothing

DMN to FPN

  • Nothing

Stress and Enrichment Factor Analyses

To examine whether the age x SES interactions in visual areas are accounted for by stress or cognitive enrichment, I first sought to characterize these two.

Stress

Per Allyson on 8/9, used ACES sum score, PSS sum score, and WLBQ individual items for stress, along with the first two items on the Neighborhood questionnaire. A factor analysis of stress variables pulled from the PSS, ACES sum, WLBQ, and the first two items on the neighborhood questionnaire revealed that we should extract 2 (or 4) factors, depending on which scree plot you look at.

I extracted the second factor in the factor analysis, which loaded more heavily on child ACES and PSS sum than the first factor, which was dominated by the WLBQ questions. Regardless of which factor I take, though, or which package I use, the results do not change.

I examined whether it accounted for the age x SES composite effect on visual-DAN networks or DMN-VAN connectivity. It does not. However, note that there is slightly less power to detect an interaction with 59 as opposed to 65 subjects who have all data for the stress factor, but when examining only that subset, the age x SES composite on visual networks still holds, even when controlling for the first factor of stress.

I’m not completely sure about the correctness of the approach I took, but did try two different factor analysis packages, extracted the second or first factor (which was slightly different) using both, and examined it here, and get similar results.

Also, this factor does not significantly predict any main effects on any specific networks after FDR correction across networks, or any whole-brain segregation measures. It is negatively correlated with SES (!), but positively correlated with PSS and child ACES (obviously).

However, when controlling for race and ethnicity, child ACES sum is marginally interacting with age to be predictive of visual-DAN connectivity, even when controlling for SES. The same is true for predicting DMN-DAN connectivity.

Cognitive Enrichment

I used the HOME subscale scores, as well as the literacy and numeracy questionnaire question about the number of child books and the number of hours per week the child was read to. I also included literacy and numeracy subscales for frequency of other activities at home: Music, Visual Arts, Pretend Play, Spatial, Fine Motor.

Examining a scree plot yields that either 3 or 2 factors should be extracted. N=59 have information on cognitive stimulation.

We see that when including the first factor of cognitive stimulation, which loads mostly on the literacy and numeracy questionnaire, and less on the HOME or the number of child books, it partially accounts for the effect of SES on visual- DAN connectivity. The age x cognitive stimulation interaction is significant, however, including the first factor in the age* ses_composite model does not affect the strength of the age x SES interaction predicting VIS-DAN connectivity. Cognitive stimulation is not predictive of DMN-VAN connectivity.

It is negatively correlated with SES, strongly so (!), and slightly positively correlated with child ACES! Odd. Looks like in our sample, HOME is weakly positively correlated with SES, but Litnum is even more stronly negatively correlated with SES.

Also, this factor does not significantly predict any main effects on any specific networks after FDR correction across networks, or any whole-brain segregation measures or interactions with age.

### Questions from Allyson about Stress & Cog Stim Analyses

I’m wondering if it’s better to just use the HOME, and maybe just the items from the HOME that are related to SES (or the items that are not endorsed by 99% of families- I think that ends up being the same item set).

Which variables are these? I decided to not pull specifically variables that relate to SES, that seems like double-dipping/fishing for an effect, so choose an SD cutoff based on looking at the variance across different questions.

The sum of variables on the HOME that have an SD > 0.2 is correlated positively with SES, at Kendall’s tau= 0.1909273

We examine whether this HOME sum total accounts for the age x SES interaction in predicting VIS-DAN connectivity. It does not, and controlling for it does not affect the significance of the age x SES interaction in predicting connectivity. It also does not show a main effect on VIS-DAN connectivity.

Also, this HOME sum does not significantly any whole-brain segregation measures or interactions with age. It also does not predict any main effects on any specific networks after FDR correction across networks (except a p=0.02 main effect on DAN-LIM, but I’ve done so many tests now on the stress-cog stim variables, I’m not sure this is anything.).

It might be worth confirming that the individual stress measures don’t show an ageXstress interaction.

Neither the summed PSS, the 3 category ACES, the summmed WLBQ, or neighborhood perceived quality (which include those safety questions) show main effects or interactions with age predicting VIS-DAN connectivity. However, including race in the model does make some of these stress measures marginally predictive of VIS-DAN or DMN-VAN connectivity.

Questions, Notes, & Future Directions

What do age and age x SES effects look like in only the oldest kids?

  • I also examined whether this is being driven by the younger or the older kids. We know already that the DMN-VAN age decrease does not hold if excluding the oldest kids. When looking only at kids over age 6 (n=49), DMN-DAN and DMN-VAN age effects still survive FDR correction, as does decreasing LIM-FPN connectivity with age. No SES main effects in these networks, and the only SES main effect that survives FDR correction is SMN-FPN (as in the larger sample). Age x SES interactions are not significant (though VIS-DAN and DMN-DMN are p=0.15), with the exception of FPN-DMN connectivity, which shows a significant age x SES effect such that higher-SES children show a strong association between decreasing FPN-DMN connectivity and age than lower-SES children.

Run 1 vs Run 2

When using run 1 instead of run 2 for the n=26 participants with two runs in the main pipeline used here, the age effects still hold, but do not pass FDR correction (they’re marginal, p=0.11). Age x SES interaction on visual-DAN networks still holds, but age x SES interaction on DMN-DAN networks is p=0.2.

Replication parcellation: Schaefer200

Age and measures of network segregation

We used Schaefer200 as the replication parcellation, since it has nice correspondence to Schaefer400 and the Yeo7 systems.

All previous findings significant increases in network segregation with age hold in our replication parcellation! The only finding that is inconsistent is the number of communities detected using modul_avgarity maximization decreasing with age, which was likely spurious anyways.

As a reminder, we controlled for sex, mean framewise displacement, percent of FD spikes in the data, the number of volumes a participant had, and the average weight of the functional network.

Also, no non-linear effects of age.

Environment and Measures of Network Segregation

However, as above, there are effects of the binned ACES (into 0-1, 2, or 3+) on within- and between-network connectivity as well as the participation coefficient. We see that 2 ACES is associated with significantly less within- and more between-network connectivity (p=0.035906), and a higher average participation coefficient (p=0.0550084), lower system segregation (p=0.051243).

System-level effects

Again, we find the strongest effects in Vis-DAN, DMN-DAN, and DMN-VAN, which are the only effects that pass fdr correction.

All three effects survive FDR correction across the tests conducted, with DMN-VAN p_fdr=0.0028636, DMN-DAN p_fdr= 0.0358089, Vis-DAN p_fdr= 0.0340196.

System-specific connectivity and environment main effects

Again, the effects shown above replicate.

Visual to DAN

  • Significant + effect of parent education on avg connectivity (p = 0.01), not significant controlling for race and ethnicity.

DMN to DAN

  • Significant - effect of ses composite on avg connectivity (p =0.01), not significant controlling for race and ethnicity.
  • Significant - effect of parent education on avg connectivity (p =0.02), not significant controlling for race and ethnicity.
  • Significant - effect of median income on avg connectivity (p =0.02), not significant controlling for race and ethnicity.

DMN to DAN

  • Nothing

All systems connectivity and environment main effects

As Allyson suggested on 8/6, I ran models instead looking for a main effect of environment across all networks and fdr-corrected in the replication parcellation as well. I chose the SES composite for this analysis, since that seemed the most principled. No networks showed significant or marginal main effects of the environment after FDR correction. When controlling for race and ethnicity, however, networks that showed significant effects of the environment were MOT-FPN, p_FDR=0.0381125 and VAN-LIM p_fdr=0.1557001.

System-specific connectivity and environment interactions

Visual to DAN

  • Significant interaction between age and ses composite on avg connectivity (p = 0.0008)
  • Significant interaction between age and income on avg connectivity (p = 0.01)
  • Significant interaction between age and parent ed on avg connectivity (p = 0.01)

DMN to DAN

  • Significant interaction between age and parent ed on avg connectivity (p = 0.03)

DMN to VAN

  • Significant interaction between age and ses composite on avg connectivity (p = 0.04)
  • Marginal interaction between age and income on avg connectivity (p = 0.06)
  • Marginal interaction between age and parent ed on avg connectivity (p = 0.09)

As suggested by Allyson on 8/6, I also ran some models looking at whether VIS-VIS, DMN-DMN, or FPN-DMN show age x SES interactions.

Vis to Vis

  • Nothing

DMN to DMN

  • Nothing

DMN to FPN

  • Nothing

Preprocessing Pipelines & Robustness

I examined whether the results above hold with different preprocessing pipelines, and to what extent the choice of pipeline affects how much data we have to work with.

How much data do we have/lose?

At the moment, my feeling is that we need something like 120-150 volumes of good data per participant to say that we could extract good resting-state networks from them. I think 120 is low but defensible.

See below for histograms of the number of un-spike-regressed volumes we have in Run 1 (or the first run) only with different pipelines. QA criteria are, as delineated above, anyone with > 50% frames censored (following Yeo papers) or mean motion > 1 mm, or max motion over 10 mm. These histograms include people who didn’t pass QA criteria, so as not to obscur the distribution.

Using both runs

What if we add the two runs for subjects who have more than one, what does that buy us?

Not much at all, it turns out. Looks like most of the kids we’re losing to high motion are those that were scanned before we started doing two runs of rest. Also, just as an aside, I do think we’ve gotten better at scanning than we used to be, but that doesn’t hold up quantitatively.

Pipeline